> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getnao.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Settings

> Tune per-model inference parameters and the Python execution limit

Admins can tune how the agent calls each model, and how long it is allowed to run Python. Both live in the admin panel under **Settings** -> **Project**.

## Model parameters

Each enabled model can carry its own inference settings, stored per project and per provider. Models without overrides use the provider's defaults.

1. Go to **Settings** -> **Project** -> **Models**.
2. Add or edit a provider.
3. In **Enabled Models**, click the **...** button on an enabled model to open the **Model parameters** dialog.
4. Set the parameters you want and click **Save**.

Custom models you added yourself have the same edit button. Clearing every field removes the overrides for that model.

### Available parameters

The dialog only shows the controls the selected model actually accepts, so the list changes from model to model:

| Parameter                             | Applies to                                                                                                                                                                                                                                                                                                                                                                 |
| ------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Thinking effort**                   | Models with adaptive thinking (recent Claude, GPT-5.x, Gemini 3.x, OpenRouter, Bedrock Claude 4.6+). The available levels depend on the model.                                                                                                                                                                                                                             |
| **Thinking budget (tokens)**          | Models with budget-based extended thinking (Claude 4.5 era, Gemini 2.5, legacy Bedrock Claude).                                                                                                                                                                                                                                                                            |
| **Temperature**, **Top P**, **Top K** | Sampling models. Upper bounds are model-specific, and **Top K** only appears where the model supports it.                                                                                                                                                                                                                                                                  |
| **Max output tokens**                 | Most models.                                                                                                                                                                                                                                                                                                                                                               |
| Provider extras                       | Provider-specific controls, for example **Parallel tool calls**, **Send reasoning back**, **Speed** and **Inference geography** on Anthropic, **Text verbosity**, **Reasoning summary**, **Max tool calls** and **Service tier** on OpenAI, **Include thoughts**, **Safety threshold** and **Media resolution** on Google, **Safe prompt** and document limits on Mistral. |

<Note>
  On Claude models (direct, or through Vertex or Bedrock), **Top P** and **Temperature** are mutually exclusive - when both are set, nao sends only the temperature. Values outside a model's accepted range are clamped before the call.
</Note>

Thinking effort and max output tokens directly change token spend. See [Budgets](/nao-agent/chat/admin/budgets) to cap that spend per provider.

## Python execution duration

When the Python sandbox is enabled, you can cap how long a single Python execution may run.

1. Go to **Settings** -> **Project** -> **Agent**.
2. In the **Experimental** card, set **Python execution duration**.

The value is in seconds, defaults to **30**, and must be between **1** and **600**. Code that runs longer is stopped.